Nonparametric Tests – Resolving Discrepancies Between Omnibus and Post Hoc Tests

dunn-testhypothesis testingkruskal-wallis test”nonparametricr

So I am running a simple Dunn's Test on my data. After running the Kruskal-Wallis H Test, it returned a result


data:  Mean_Lesion_Height by ISOLATE
Kruskal-Wallis chi-squared = 56.324, df = 40, p-value = 0.04499

Clearly, the p value is <0.05 suggesting that there is difference in the median of the groups I tested.
To identify which of these groups showed the significant difference. I run a post hoc test of Dunn's Test using dunnTest() and gave me the result

 Kruskal-Wallis rank sum test 
  
 data: x and g 
 Kruskal-Wallis chi-squared = 56.3235, df = 40, p-value = 0.05

Suddenly, p-value changed and dunnTest() can no longer identify which groups are different.

to perform the Dunn's Test, I set the method to "bonferroni", I also tried "none", "sidak", "holm", "hs", hochberg", "bh", "by", and still get a p-value of 0.05 in dunnTest().

Can somebody please explain to me why this is happening? Also, how do I choose the correct p adjustment method?

Best Answer

I only see one contrast group in the post-hoc test. Make sure that your IV is coded as a factor.

It can happen that the global test is significant while none of the post-hoc comparisons is due to p-value adjustment.

Also strictly speaking the Kruskal-Wallis test just suggests that the distributions differ. We can only assume that medians differ based on that.

Related Question